File size: 12,788 Bytes
fbad523 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 | import React, { useState } from 'react';
import { Search, Zap, Layers, Sparkles } from 'lucide-react';
export const MathInspectorTab: React.FC = () => {
const [selectedParadigm, setSelectedParadigm] = useState<'sparse' | 'dense' | 'learned-sparse' | 'cross-encoder'>('dense');
return (
<div className="space-y-5 max-w-5xl">
{/* Overview Card */}
<div className="border border-zinc-800 bg-zinc-900/40 p-4 rounded-lg space-y-2.5 font-sans">
<div className="flex items-center justify-between">
<h2 className="text-base font-semibold text-zinc-100">
The Unified Inner-Product Formulation
</h2>
<span className="text-[11px] font-mono text-zinc-500">
Lin et al. (2021)
</span>
</div>
<p className="text-xs text-zinc-400 leading-relaxed">
Modern information retrieval maps both lexical and neural models to a unified mathematical objective:
<strong> Maximum Inner Product Search (MIPS)</strong> over vector spaces.
Sparse, dense, and learned sparse methods differ primarily in dimensionality, sparsity, and indexing data structures.
</p>
{/* Math Equation */}
<div className="p-3 bg-zinc-950 border border-zinc-850 rounded font-mono text-center text-xs text-zinc-200 my-1">
Score(q, d) = ⟨ E<sub>Q</sub>(q), E<sub>D</sub>(d) ⟩ = ∑<sub>i</sub> E<sub>Q</sub>(q)<sub>i</sub> · E<sub>D</sub>(d)<sub>i</sub>
</div>
</div>
{/* Paradigm Tabs */}
<div className="grid grid-cols-2 md:grid-cols-4 gap-2 text-xs">
<button
onClick={() => setSelectedParadigm('sparse')}
className={`p-3 rounded-lg border text-left transition flex items-start space-x-2 cursor-pointer ${
selectedParadigm === 'sparse'
? 'bg-amber-100 dark:bg-amber-500/10 border-amber-400 dark:border-amber-500/50 text-amber-950 dark:text-amber-200 ring-1 ring-amber-400/40 font-semibold dark:font-medium'
: 'bg-zinc-100/60 dark:bg-zinc-950 border-zinc-300 dark:border-zinc-850 text-zinc-600 dark:text-zinc-400 hover:text-zinc-900 dark:hover:text-zinc-200 hover:bg-zinc-200/60 dark:hover:bg-zinc-900'
}`}
>
<Search className={`w-4 h-4 mt-0.5 shrink-0 ${selectedParadigm === 'sparse' ? 'text-amber-700 dark:text-amber-400' : 'text-zinc-400 dark:text-zinc-500'}`} />
<div>
<div className="font-semibold">Unsupervised Sparse</div>
<div className="text-[11px] text-zinc-500 mt-0.5">BM25 / Lucene Postings</div>
</div>
</button>
<button
onClick={() => setSelectedParadigm('dense')}
className={`p-3 rounded-lg border text-left transition flex items-start space-x-2 cursor-pointer ${
selectedParadigm === 'dense'
? 'bg-cyan-100 dark:bg-cyan-500/10 border-cyan-400 dark:border-cyan-500/50 text-cyan-950 dark:text-cyan-200 ring-1 ring-cyan-400/40 font-semibold dark:font-medium'
: 'bg-zinc-100/60 dark:bg-zinc-950 border-zinc-300 dark:border-zinc-850 text-zinc-600 dark:text-zinc-400 hover:text-zinc-900 dark:hover:text-zinc-200 hover:bg-zinc-200/60 dark:hover:bg-zinc-900'
}`}
>
<Zap className={`w-4 h-4 mt-0.5 shrink-0 ${selectedParadigm === 'dense' ? 'text-cyan-700 dark:text-cyan-400' : 'text-zinc-400 dark:text-zinc-500'}`} />
<div>
<div className="font-semibold">Supervised Dense</div>
<div className="text-[11px] text-zinc-500 mt-0.5">BGE / Faiss HNSW</div>
</div>
</button>
<button
onClick={() => setSelectedParadigm('learned-sparse')}
className={`p-3 rounded-lg border text-left transition flex items-start space-x-2 cursor-pointer ${
selectedParadigm === 'learned-sparse'
? 'bg-purple-100 dark:bg-purple-500/10 border-purple-400 dark:border-purple-500/50 text-purple-950 dark:text-purple-200 ring-1 ring-purple-400/40 font-semibold dark:font-medium'
: 'bg-zinc-100/60 dark:bg-zinc-950 border-zinc-300 dark:border-zinc-850 text-zinc-600 dark:text-zinc-400 hover:text-zinc-900 dark:hover:text-zinc-200 hover:bg-zinc-200/60 dark:hover:bg-zinc-900'
}`}
>
<Layers className={`w-4 h-4 mt-0.5 shrink-0 ${selectedParadigm === 'learned-sparse' ? 'text-purple-700 dark:text-purple-400' : 'text-zinc-400 dark:text-zinc-500'}`} />
<div>
<div className="font-semibold">Learned Sparse</div>
<div className="text-[11px] text-zinc-500 mt-0.5">SPLADE / Lucene Impact</div>
</div>
</button>
<button
onClick={() => setSelectedParadigm('cross-encoder')}
className={`p-3 rounded-lg border text-left transition flex items-start space-x-2 cursor-pointer ${
selectedParadigm === 'cross-encoder'
? 'bg-emerald-100 dark:bg-emerald-500/10 border-emerald-400 dark:border-emerald-500/50 text-emerald-950 dark:text-emerald-200 ring-1 ring-emerald-400/40 font-semibold dark:font-medium'
: 'bg-zinc-100/60 dark:bg-zinc-950 border-zinc-300 dark:border-zinc-850 text-zinc-600 dark:text-zinc-400 hover:text-zinc-900 dark:hover:text-zinc-200 hover:bg-zinc-200/60 dark:hover:bg-zinc-900'
}`}
>
<Sparkles className={`w-4 h-4 mt-0.5 shrink-0 ${selectedParadigm === 'cross-encoder' ? 'text-emerald-700 dark:text-emerald-400' : 'text-zinc-400 dark:text-zinc-500'}`} />
<div>
<div className="font-semibold">Cross-Encoder</div>
<div className="text-[11px] text-zinc-500 mt-0.5">RankLLM / Full Attention</div>
</div>
</button>
</div>
{/* Details Box */}
<div className="border border-zinc-800 bg-zinc-900/30 p-4.5 rounded-lg space-y-3.5 text-xs font-sans">
{selectedParadigm === 'sparse' && (
<div className="space-y-3">
<h3 className="text-sm font-semibold text-amber-200">
Unsupervised Sparse Retrieval (BM25)
</h3>
<div className="grid grid-cols-1 md:grid-cols-3 gap-2 font-mono text-[11px]">
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Dimensionality:</span>
<span className="text-zinc-200 font-medium">|V| ≈ 30,000+ words</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Sparsity:</span>
<span className="text-zinc-200 font-medium">> 99.9% zeros</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Index Structure:</span>
<span className="text-zinc-200 font-medium">Inverted Posting Lists (Lucene)</span>
</div>
</div>
<p className="text-zinc-400 leading-relaxed">
In BM25, query and document vectors lie in high-dimensional lexical vocabulary space.
Only non-zero entries correspond to tokens present in the document.
The dot product is evaluated via inverted posting lists without scanning non-matching documents.
</p>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850 font-mono text-[11px] text-amber-200/90">
E_Q(q)_i = 1 (if token i ∈ q, else 0)<br />
E_D(d)_i = IDF(i) · [ TF(i,d)·(k₁+1) / (TF(i,d) + k₁·(1 - b + b·|D|/avgdl)) ]
</div>
</div>
)}
{selectedParadigm === 'dense' && (
<div className="space-y-3">
<h3 className="text-sm font-semibold text-cyan-200">
Supervised Dense Retrieval (Bi-Encoder)
</h3>
<div className="grid grid-cols-1 md:grid-cols-3 gap-2 font-mono text-[11px]">
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Dimensionality:</span>
<span className="text-zinc-200 font-medium">d = 768 continuous floats</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Sparsity:</span>
<span className="text-zinc-200 font-medium">0% zeros (fully dense)</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Index Structure:</span>
<span className="text-zinc-200 font-medium">HNSW Proximity Graph (Faiss)</span>
</div>
</div>
<p className="text-zinc-400 leading-relaxed">
Dense retrieval projects text into low-dimensional semantic vector spaces via dual transformers:
<code className="text-cyan-300 font-mono ml-1">E_Q(q) = Transformer_Q(q)</code> and
<code className="text-cyan-300 font-mono ml-1">E_D(d) = Transformer_D(d)</code>.
Because all coordinates are non-zero, inverted lists cannot be used. Instead, approximate nearest neighbor (ANN) search algorithms like
HNSW traverse multi-layer proximity graphs in logarithmic time.
</p>
</div>
)}
{selectedParadigm === 'learned-sparse' && (
<div className="space-y-3">
<h3 className="text-sm font-semibold text-purple-950 dark:text-purple-200">
Learned Sparse Representations (SPLADE)
</h3>
<div className="grid grid-cols-1 md:grid-cols-3 gap-2 font-mono text-[11px]">
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Dimensionality:</span>
<span className="text-zinc-200 font-medium">|V| = 30,522 (WordPiece Vocab)</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Sparsity:</span>
<span className="text-zinc-200 font-medium">~95% zeros (regularized)</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Index Structure:</span>
<span className="text-zinc-200 font-medium">Lucene Impact Index</span>
</div>
</div>
<p className="text-zinc-400 leading-relaxed">
SPLADE combines the efficiency of inverted indexes with deep contextual vocabulary expansion.
A masked language model predicts non-negative weights over the full vocabulary, expanding terms with synonyms and related concepts.
In Anserini, floating-point weights are quantized to integers and indexed directly in Lucene as synthetic term frequencies.
</p>
</div>
)}
{selectedParadigm === 'cross-encoder' && (
<div className="space-y-3">
<h3 className="text-sm font-semibold text-emerald-200">
Cross-Encoder Listwise Reranking (RankLLM / monoT5)
</h3>
<div className="grid grid-cols-1 md:grid-cols-3 gap-2 font-mono text-[11px]">
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Architecture:</span>
<span className="text-zinc-200 font-medium">Joint Full Attention (q × d)</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Complexity:</span>
<span className="text-zinc-200 font-medium">O((|q| + |d|)² · L) per document</span>
</div>
<div className="p-2.5 bg-zinc-950 rounded border border-zinc-850">
<span className="text-zinc-500 block">Index Structure:</span>
<span className="text-zinc-200 font-medium">None (2nd-stage reranker on Top-K)</span>
</div>
</div>
<p className="text-zinc-400 leading-relaxed">
Unlike bi-encoders which compute independent representations, cross-encoders concatenate query and passage
into a single transformer context. Every token attends directly to every other token, capturing nuanced nuances,
negation, and syntactic dependencies at higher inference cost.
</p>
</div>
)}
</div>
</div>
);
};
|